dimethenamid_2018.rmd

---
title: Example evaluations of the dimethenamid data from 2018
author: Johannes Ranke
date: Last change 1 July 2022, built on `r Sys.setlocale("LC_TIME", "C"); format(Sys.Date(), format = "%d %b %Y")`
output:
  html_document:
    toc: true
    toc_float: true
    code_folding: hide
    fig_retina: null
bibliography: ../references.bib
vignette: >
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

[Wissenschaftlicher Berater, Kronacher Str. 12, 79639 Grenzach-Wyhlen, Germany](http://www.jrwb.de)

```{r, include = FALSE}
require(knitr)
require(mkin)
require(nlme)
options(digits = 5)
opts_chunk$set(
  comment = "",
  tidy = FALSE,
  cache = TRUE
)
```

# Introduction

A first analysis of the data analysed here was presented in a recent journal
article on nonlinear mixed-effects models in degradation kinetics [@ranke2021].
That analysis was based on the `nlme` package and a development version
of the `saemix` package that was unpublished at the time. Meanwhile, version
3.0 of the `saemix` package is available from the CRAN repository. Also, it
turned out that there was an error in the handling of the Borstel data in the
mkin package at the time, leading to the duplication of a few data points from
that soil. The dataset in the mkin package has been corrected, and the interface
to `saemix` in the mkin package has been updated to use the released version.

This vignette is intended to present an up to date analysis of the data, using the
corrected dataset and released versions of `mkin` and `saemix`.

# Data

Residue data forming the basis for the endpoints derived in the conclusion on
the peer review of the pesticide risk assessment of dimethenamid-P published by
the European Food Safety Authority (EFSA) in 2018 [@efsa_2018_dimethenamid]
were transcribed from the risk assessment report [@dimethenamid_rar_2018_b8]
which can be downloaded from the Open EFSA repository
[https://open.efsa.europa.eu/study-inventory/EFSA-Q-2014-00716](https://open.efsa.europa.eu).

The data are [available in the mkin
package](https://pkgdown.jrwb.de/mkin/reference/dimethenamid_2018.html). The
following code (hidden by default, please use the button to the right to show
it) treats the data available for the racemic mixture dimethenamid (DMTA) and
its enantiomer dimethenamid-P (DMTAP) in the same way, as no difference between
their degradation behaviour was identified in the EU risk assessment. The
observation times of each dataset are multiplied with the corresponding
normalisation factor also available in the dataset, in order to make it
possible to describe all datasets with a single set of parameters.

Also, datasets observed in the same soil are merged, resulting in dimethenamid
(DMTA) data from six soils.

```{r dimethenamid_data}
library(mkin, quietly = TRUE)
dmta_ds <- lapply(1:7, function(i) {
  ds_i <- dimethenamid_2018$ds[[i]]$data
  ds_i[ds_i$name == "DMTAP", "name"] <-  "DMTA"
  ds_i$time <- ds_i$time * dimethenamid_2018$f_time_norm[i]
  ds_i
})
names(dmta_ds) <- sapply(dimethenamid_2018$ds, function(ds) ds$title)
dmta_ds[["Elliot"]] <- rbind(dmta_ds[["Elliot 1"]], dmta_ds[["Elliot 2"]])
dmta_ds[["Elliot 1"]] <- NULL
dmta_ds[["Elliot 2"]] <- NULL
```

# Parent degradation

We evaluate the observed degradation of the parent compound using simple
exponential decline (SFO) and biexponential decline (DFOP), using constant
variance (const) and a two-component variance (tc) as error models.

## Separate evaluations

As a first step, to get a visual impression of the fit of the different models,
we do separate evaluations for each soil using the mmkin function from the
mkin package:

```{r f_parent_mkin}
f_parent_mkin_const <- mmkin(c("SFO", "DFOP"), dmta_ds,
  error_model = "const", quiet = TRUE)
f_parent_mkin_tc <- mmkin(c("SFO", "DFOP"), dmta_ds,
  error_model = "tc", quiet = TRUE)
```

The plot of the individual SFO fits shown below suggests that at least in some
datasets the degradation slows down towards later time points, and that the
scatter of the residuals error is smaller for smaller values (panel to the
right):

```{r f_parent_mkin_sfo_const}
plot(mixed(f_parent_mkin_const["SFO", ]))
```

Using biexponential decline (DFOP) results in a slightly more random
scatter of the residuals:

```{r f_parent_mkin_dfop_const}
plot(mixed(f_parent_mkin_const["DFOP", ]))
```

The population curve (bold line) in the above plot results from taking the mean
of the individual transformed parameters, i.e. of log k1 and log k2, as well as
of the logit of the g parameter of the DFOP model). Here, this procedure
does not result in parameters that represent the degradation well, because in some
datasets the fitted value for k2 is extremely close to zero, leading to a log
k2 value that dominates the average. This is alleviated if only rate constants
that pass the t-test for significant difference from zero (on the untransformed
scale) are considered in the averaging:

```{r f_parent_mkin_dfop_const_test}
plot(mixed(f_parent_mkin_const["DFOP", ]), test_log_parms = TRUE)
```

While this is visually much more satisfactory, such an average procedure could
introduce a bias, as not all results from the individual fits enter the
population curve with the same weight. This is where nonlinear mixed-effects
models can help out by treating all datasets with equally by fitting a
parameter distribution model together with the degradation model and the error
model (see below).

The remaining trend of the residuals to be higher for higher predicted residues
is reduced by using the two-component error model:

```{r f_parent_mkin_dfop_tc_test}
plot(mixed(f_parent_mkin_tc["DFOP", ]), test_log_parms = TRUE)
```

However, note that in the case of using this error model, the fits to the
Flaach and BBA 2.3 datasets appear to be ill-defined, indicated by the fact
that they did not converge:

```{r f_parent_mkin_dfop_tc_print}
print(f_parent_mkin_tc["DFOP", ])
```

## Nonlinear mixed-effects models

Instead of taking a model selection decision for each of the individual fits, we fit
nonlinear mixed-effects models (using different fitting algorithms as implemented in
different packages) and do model selection using all available data at the same time.
In order to make sure that these decisions are not unduly influenced by the
type of algorithm used, by implementation details or by the use of wrong control
parameters, we compare the model selection results obtained with different R
packages, with different algorithms and checking control parameters.

### nlme

The nlme package was the first R extension providing facilities to fit nonlinear
mixed-effects models. We would like to do model selection from all four
combinations of degradation models and error models based on the AIC.
However, fitting the DFOP model with constant variance and using default
control parameters results in an error, signalling that the maximum number
of 50 iterations was reached, potentially indicating overparameterisation.
Nevertheless, the algorithm converges when the two-component error model is
used in combination with the DFOP model. This can be explained by the fact
that the smaller residues observed at later sampling times get more
weight when using the two-component error model which will counteract the
tendency of the algorithm to try parameter combinations unsuitable for
fitting these data.

```{r f_parent_nlme, warning = FALSE}
library(nlme)
f_parent_nlme_sfo_const <- nlme(f_parent_mkin_const["SFO", ])
# f_parent_nlme_dfop_const <- nlme(f_parent_mkin_const["DFOP", ])
f_parent_nlme_sfo_tc <- nlme(f_parent_mkin_tc["SFO", ])
f_parent_nlme_dfop_tc <- nlme(f_parent_mkin_tc["DFOP", ])
```

Note that a certain degree of overparameterisation is also indicated by a
warning obtained when fitting DFOP with the two-component error model ('false
convergence' in the 'LME step' in iteration 3). However, as this warning does
not occur in later iterations, and specifically not in the last of the
`r f_parent_nlme_dfop_tc$numIter` iterations, we can ignore this warning.

The model comparison function of the nlme package can directly be applied
to these fits showing a much lower AIC for the DFOP model fitted with the
two-component error model. Also, the likelihood ratio test indicates that this
difference is significant as the p-value is below 0.0001.

```{r AIC_parent_nlme}
anova(
  f_parent_nlme_sfo_const, f_parent_nlme_sfo_tc, f_parent_nlme_dfop_tc
)
```

In addition to these fits, attempts were also made to include correlations
between random effects by using the log Cholesky parameterisation of the matrix
specifying them. The code used for these attempts can be made visible below.

```{r f_parent_nlme_logchol, warning = FALSE, eval = FALSE}
f_parent_nlme_sfo_const_logchol <- nlme(f_parent_mkin_const["SFO", ],
  random = nlme::pdLogChol(list(DMTA_0 ~ 1, log_k_DMTA ~ 1)))
anova(f_parent_nlme_sfo_const, f_parent_nlme_sfo_const_logchol)
f_parent_nlme_sfo_tc_logchol <- nlme(f_parent_mkin_tc["SFO", ],
  random = nlme::pdLogChol(list(DMTA_0 ~ 1, log_k_DMTA ~ 1)))
anova(f_parent_nlme_sfo_tc, f_parent_nlme_sfo_tc_logchol)
f_parent_nlme_dfop_tc_logchol <- nlme(f_parent_mkin_const["DFOP", ],
  random = nlme::pdLogChol(list(DMTA_0 ~ 1, log_k1 ~ 1, log_k2 ~ 1, g_qlogis ~ 1)))
anova(f_parent_nlme_dfop_tc, f_parent_nlme_dfop_tc_logchol)
```

While the SFO variants converge fast, the additional parameters introduced
by this lead to convergence warnings for the DFOP model. The model comparison
clearly show that adding correlations between random effects does not improve
the fits.

The selected model (DFOP with two-component error) fitted to the data assuming
no correlations between random effects is shown below.

```{r plot_parent_nlme}
plot(f_parent_nlme_dfop_tc)
```

### saemix

The saemix package provided the first Open Source implementation of the
Stochastic Approximation to the Expectation Maximisation (SAEM) algorithm.
SAEM fits of degradation models can be conveniently performed using an
interface to the saemix package available in current development versions of
the mkin package.

The corresponding SAEM fits of the four combinations of degradation and error
models are fitted below. As there is no convergence criterion implemented in
the saemix package, the convergence plots need to be manually checked for every
fit. We define control settings that work well for all the parent data fits
shown in this vignette.

```{r saemix_control, results='hide'}
library(saemix)
saemix_control <- saemixControl(nbiter.saemix = c(800, 300), nb.chains = 15,
    print = FALSE, save = FALSE, save.graphs = FALSE, displayProgress = FALSE)
saemix_control_moreiter <- saemixControl(nbiter.saemix = c(1600, 300), nb.chains = 15,
    print = FALSE, save = FALSE, save.graphs = FALSE, displayProgress = FALSE)
saemix_control_10k <- saemixControl(nbiter.saemix = c(10000, 300), nb.chains = 15,
    print = FALSE, save = FALSE, save.graphs = FALSE, displayProgress = FALSE)
```

The convergence plot for the SFO model using constant variance is shown below.

```{r f_parent_saemix_sfo_const, results = 'hide', dependson = "saemix_control"}
f_parent_saemix_sfo_const <- mkin::saem(f_parent_mkin_const["SFO", ], quiet = TRUE,
  control = saemix_control, transformations = "saemix")
plot(f_parent_saemix_sfo_const$so, plot.type = "convergence")
```

Obviously the selected number of iterations is sufficient to reach convergence.
This can also be said for the SFO fit using the two-component error model.

```{r f_parent_saemix_sfo_tc, results = 'hide', dependson = "saemix_control"}
f_parent_saemix_sfo_tc <- mkin::saem(f_parent_mkin_tc["SFO", ], quiet = TRUE,
  control = saemix_control, transformations = "saemix")
plot(f_parent_saemix_sfo_tc$so, plot.type = "convergence")
```

When fitting the DFOP model with constant variance (see below), parameter
convergence is not as unambiguous.

```{r f_parent_saemix_dfop_const, results = 'show', dependson = "saemix_control"}
f_parent_saemix_dfop_const <- mkin::saem(f_parent_mkin_const["DFOP", ], quiet = TRUE,
  control = saemix_control, transformations = "saemix")
plot(f_parent_saemix_dfop_const$so, plot.type = "convergence")
print(f_parent_saemix_dfop_const)
```

While the other parameters converge to credible values, the variance of k2
(`omega2.k2`) converges to a very small value. The printout of the `saem.mmkin`
model shows that the estimated standard deviation of k2 across the population
of soils (`SD.k2`) is ill-defined, indicating overparameterisation of this model.

When the DFOP model is fitted with the two-component error model, we also
observe that the estimated variance of k2 becomes very small, while being
ill-defined, as illustrated by the excessive confidence interval of `SD.k2`.

```{r f_parent_saemix_dfop_tc, results = 'show', dependson = "saemix_control"}
f_parent_saemix_dfop_tc <- mkin::saem(f_parent_mkin_tc["DFOP", ], quiet = TRUE,
  control = saemix_control, transformations = "saemix")
f_parent_saemix_dfop_tc_moreiter <- mkin::saem(f_parent_mkin_tc["DFOP", ], quiet = TRUE,
  control = saemix_control_moreiter, transformations = "saemix")
plot(f_parent_saemix_dfop_tc$so, plot.type = "convergence")
print(f_parent_saemix_dfop_tc)
```

Doubling the number of iterations in the first phase of the algorithm
leads to a slightly lower likelihood, and therefore to slightly higher AIC and BIC values.
With even more iterations, the algorithm stops with an error message. This is
related to the variance of k2 approximating zero and has been submitted
as a [bug to the saemix package](https://github.com/saemixdevelopment/saemixextension/issues/29),
as the algorithm does not converge in this case.

An alternative way to fit DFOP in combination with the two-component error model
is to use the model formulation with transformed parameters as used per default
in mkin. When using this option, convergence is slower, but eventually
the algorithm stops as well with the same error message.

The four combinations (SFO/const, SFO/tc, DFOP/const and DFOP/tc) and
the version with increased iterations can be compared using the model
comparison function of the saemix package:

```{r AIC_parent_saemix, cache = FALSE}
AIC_parent_saemix <- saemix::compare.saemix(
  f_parent_saemix_sfo_const$so,
  f_parent_saemix_sfo_tc$so,
  f_parent_saemix_dfop_const$so,
  f_parent_saemix_dfop_tc$so,
  f_parent_saemix_dfop_tc_moreiter$so)
rownames(AIC_parent_saemix) <- c(
  "SFO const", "SFO tc", "DFOP const", "DFOP tc", "DFOP tc more iterations")
print(AIC_parent_saemix)
```

In order to check the influence of the likelihood calculation algorithms
implemented in saemix, the likelihood from Gaussian quadrature is added
to the best fit, and the AIC values obtained from the three methods
are compared.

```{r AIC_parent_saemix_methods, cache = FALSE}
f_parent_saemix_dfop_tc$so <-
  saemix::llgq.saemix(f_parent_saemix_dfop_tc$so)
AIC_parent_saemix_methods <- c(
  is = AIC(f_parent_saemix_dfop_tc$so, method = "is"),
  gq = AIC(f_parent_saemix_dfop_tc$so, method = "gq"),
  lin = AIC(f_parent_saemix_dfop_tc$so, method = "lin")
)
print(AIC_parent_saemix_methods)
```
The AIC values based on importance sampling and Gaussian quadrature are very
similar. Using linearisation is known to be less accurate, but still gives a
similar value.

In order to illustrate that the comparison of the three method
depends on the degree of convergence obtained in the fit, the same comparison
is shown below for the fit using the defaults for the number of iterations and
the number of MCMC chains.

When using OpenBlas for linear algebra, there is a large difference in the
values obtained with Gaussian quadrature, so the larger number of iterations
makes a lot of difference. When using the LAPACK version coming with Debian
Bullseye, the AIC based on Gaussian quadrature is almost the same as the one obtained
with the other methods, also when using defaults for the fit.

```{r AIC_parent_saemix_methods_defaults, cache = FALSE}
f_parent_saemix_dfop_tc_defaults <- mkin::saem(f_parent_mkin_tc["DFOP", ])
f_parent_saemix_dfop_tc_defaults$so <-
  saemix::llgq.saemix(f_parent_saemix_dfop_tc_defaults$so)
AIC_parent_saemix_methods_defaults <- c(
  is = AIC(f_parent_saemix_dfop_tc_defaults$so, method = "is"),
  gq = AIC(f_parent_saemix_dfop_tc_defaults$so, method = "gq"),
  lin = AIC(f_parent_saemix_dfop_tc_defaults$so, method = "lin")
)
print(AIC_parent_saemix_methods_defaults)
```

## Comparison

The following table gives the AIC values obtained with both backend packages
using the same control parameters (800 iterations burn-in, 300 iterations
second phase, 15 chains).

```{r AIC_all, cache = FALSE}
AIC_all <- data.frame(
  check.names = FALSE,
  "Degradation model" = c("SFO", "SFO", "DFOP", "DFOP"),
  "Error model" = c("const", "tc", "const", "tc"),
  nlme = c(AIC(f_parent_nlme_sfo_const), AIC(f_parent_nlme_sfo_tc), NA, AIC(f_parent_nlme_dfop_tc)),
  saemix_lin = sapply(list(f_parent_saemix_sfo_const$so, f_parent_saemix_sfo_tc$so,
    f_parent_saemix_dfop_const$so, f_parent_saemix_dfop_tc$so), AIC, method = "lin"),
  saemix_is = sapply(list(f_parent_saemix_sfo_const$so, f_parent_saemix_sfo_tc$so,
    f_parent_saemix_dfop_const$so, f_parent_saemix_dfop_tc$so), AIC, method = "is")
)
kable(AIC_all)
```

# Conclusion

A more detailed analysis of the dimethenamid dataset confirmed that the DFOP
model provides the most appropriate description of the decline of the parent
compound in these data. On the other hand, closer inspection of the results
revealed that the variability of the k2 parameter across the population of
soils is ill-defined. This coincides with the observation that this parameter
cannot robustly be quantified for some of the soils.

Regarding the regulatory use of these data, it is claimed that an improved
characterisation of the mean parameter values across the population is
obtained using the nonlinear mixed-effects models presented here. However,
attempts to quantify the variability of the slower rate constant of the
biphasic decline of dimethenamid indicate that the data are not sufficient
to characterise this variability to a satisfactory precision.

# Session Info

```{r sessionInfo, cache = FALSE}
sessionInfo()
```

# References

<!-- vim: set foldmethod=syntax: -->